Underwater Vehicle Localization by Hybridization of Indirect Kalman Filter and Neural Network

Abstract

Non-linearity in the underwater robotics system has a significant influence on navigation. The most commonly used filtering algorithm is an Extended Kalman filter or Indirect Kalman filter, which suffers from non-linear system dynamics and non-linearity in sensor measurements. In addition, the Unscented Kalman filter and the Particle filter estimation accuracy depend on particles and sigma points which contribute to high computation load. The processing load has a direct relationship with power consumption which is undesirable for long missions. The work in this paper proposes an improvement in underwater navigation and positioning of vehicles by a Multilayer perceptron neural network trained with Backpropagation. The proposed method utilizes the learning capability of a neural network and augments the noise handling ability of Kalman filtering for the state estimation which produces a new and improved underwater localization algorithm. Furthermore, the algorithm uses knowledge of state-space for neural network recursive learning that improves the overall response significantly. The algorithm is named BIKF, which has better accuracy and reliability than the Indirect Kalman filter for the prediction of underwater state variables.